[Xenomai-help] Handling interrupt - jitter

[Xenomai-help] Handling interrupt - jitter

[Bosko Radivojevic]

Hi,

I'm having (from my perspective) too big jitter on handling
interrupts. System is: ARM9, AT91SAM9260, Kernel 2.6.20.21, Adeos
1.8-03, Xenomai-2.4.2, interrupt is generated every 1ms, externally.
Interrupt source (ET1200) is connected to IRQ1 pin (PC15). I'm using
native skin from user space. Usually, handler 'reacts' after ~110us,
but, worst case delay is around 250us! Is this performance
normal/expected on my platform? Why the difference is so huge?

Is there anything I can do to get better results (with this hw:)?

Code looks like the example provided with Xenomai:

void tmng ()
{
   for (;;) {
      err = rt_intr_wait(&intr_desc,TM_INFINITE);
      if (err > 0) {
         [..]
      } else {
         [..]
      }
      rt_intr_enable (&intr_desc);
   }
}
main ()
{

  err = rt_intr_create(&intr_desc, "ET1200_sync", IRQ_NUMBER, 0);
  err = rt_task_create(&task_desc, "tmng", 0, 99, T_JOINABLE);

  rt_intr_enable (&intr_desc);

  rt_task_start(&task_desc,&tmng,NULL);
}

Thanks

Re: [Xenomai-help] Handling interrupt - jitter

[Gilles Chanteperdrix]

On Fri, Apr 18, 2008 at 4:33 PM, Bosko Radivojevic
<bosko.radivojevic@domain.hid> wrote:
> Hi,
>
>  I'm having (from my perspective) too big jitter on handling
>  interrupts. System is: ARM9, AT91SAM9260, Kernel 2.6.20.21, Adeos
>  1.8-03, Xenomai-2.4.2, interrupt is generated every 1ms, externally.

It's funny to notice how much you asked for the 2.6.24 patch for
actually not using it.

>  Interrupt source (ET1200) is connected to IRQ1 pin (PC15). I'm using
>  native skin from user space. Usually, handler 'reacts' after ~110us,
>  but, worst case delay is around 250us! Is this performance
>  normal/expected on my platform? Why the difference is so huge?

>
>  Is there anything I can do to get better results (with this hw:)?

250 us is the user-space dispatch latency you obtain with the
"latency" test on a properly loaded target. It is the time for a
user-space context switch with a full cache flush. To reduce this
latency you have to:
- handle interrupts in kernel space, so that you will be limited by
interrupt latency, not user-space scheduling latency
- use trunk so that user-space context switches involving real-time
threads do not happen with hardware interrupts off.

-- 
 Gilles

Re: [Xenomai-help] Handling interrupt - jitter

[Bosko Radivojevic]

On Fri, Apr 18, 2008 at 4:48 PM, Gilles Chanteperdrix
<gilles.chanteperdrix@xenomai.org> wrote:

>  It's funny to notice how much you asked for the 2.6.24 patch for
>  actually not using it.

:) Actually, I've tried with the 2.6.24.3, Adeos 1.9-01 & Xenomai
2.4.3 before I posted my mail, but results are the same. And, yes, I'm
using it heavily, but because of some other 'compatibility issues' I
still haven't fully migrated to the 2.6.24. Thank you very much for
migrating Adeos to 2.6.24, I really appreciate it.

>  - handle interrupts in kernel space, so that you will be limited by
>  interrupt latency, not user-space scheduling latency
>  - use trunk so that user-space context switches involving real-time
>  threads do not happen with hardware interrupts off.

Will I get anything using xenomai-trunk, and still handling interrupt
from user-space?

Thanks,
Bosko

Re: [Xenomai-help] Handling interrupt - jitter

[Gilles Chanteperdrix]

On Fri, Apr 18, 2008 at 5:08 PM, Bosko Radivojevic
<bosko.radivojevic@domain.hid> wrote:
> On Fri, Apr 18, 2008 at 4:48 PM, Gilles Chanteperdrix
>  <gilles.chanteperdrix@xenomai.org> wrote:
>
>  >  It's funny to notice how much you asked for the 2.6.24 patch for
>  >  actually not using it.
>
>  :) Actually, I've tried with the 2.6.24.3, Adeos 1.9-01 & Xenomai
>  2.4.3 before I posted my mail, but results are the same. And, yes, I'm
>  using it heavily, but because of some other 'compatibility issues' I
>  still haven't fully migrated to the 2.6.24. Thank you very much for
>  migrating Adeos to 2.6.24, I really appreciate it.
>
>
>  >  - handle interrupts in kernel space, so that you will be limited by
>  >  interrupt latency, not user-space scheduling latency
>  >  - use trunk so that user-space context switches involving real-time
>  >  threads do not happen with hardware interrupts off.
>
>  Will I get anything using xenomai-trunk, and still handling interrupt
>  from user-space?

No, the 250us switch time is due to the full cache flush which happens
at each user-space context switch on ARM (this happens because of the
ARM VIVT cache).

The only way to avoid it would be to use ucLinux. Unfortunately, the
status of ucLinux on ARM is quite fuzzy at the moment. It seems some
nommu patches have been merged in the mainstream Linux, but the kernel
still needs further patching. Which is why I have not yet tried
Xenomai over ucLinux on ARM.

Another approach is to use the FCSE extension found on some ARMs
cores, (I think ARM 926EJ-S and onwards). Using this extension imposes
some limitations (maximum number of processes and maximum RAM size for
each process). Richard Cochran posted a patch to modify Linux to use
this extension on Adeos mailing list. I had no time yet to test it,
but I think adapting Xenomai to this patch requires little work if
any.

-- 
 Gilles

Re: [Xenomai-help] Handling interrupt - jitter

[Wolfgang Grandegger]

Gilles Chanteperdrix wrote:
> On Fri, Apr 18, 2008 at 4:33 PM, Bosko Radivojevic
> <bosko.radivojevic@domain.hid> wrote:
>> Hi,
>>
>>  I'm having (from my perspective) too big jitter on handling
>>  interrupts. System is: ARM9, AT91SAM9260, Kernel 2.6.20.21, Adeos
>>  1.8-03, Xenomai-2.4.2, interrupt is generated every 1ms, externally.
> 
> It's funny to notice how much you asked for the 2.6.24 patch for
> actually not using it.
> 
>>  Interrupt source (ET1200) is connected to IRQ1 pin (PC15). I'm using
>>  native skin from user space. Usually, handler 'reacts' after ~110us,
>>  but, worst case delay is around 250us! Is this performance
>>  normal/expected on my platform? Why the difference is so huge?
> 
>>  Is there anything I can do to get better results (with this hw:)?
> 
> 250 us is the user-space dispatch latency you obtain with the
> "latency" test on a properly loaded target. It is the time for a
> user-space context switch with a full cache flush. To reduce this
> latency you have to:
> - handle interrupts in kernel space, so that you will be limited by
> interrupt latency, not user-space scheduling latency
> - use trunk so that user-space context switches involving real-time
> threads do not happen with hardware interrupts off.

I started to do some external interrupt latency measurements with
gpioirqbench showing also the effect of handling the interrupt in
user-space, kernel-space or even the interrupt handler on two PowerPC
systems. For further information, have a look to

  http://www.denx.de/wiki/DULG/AN2008_03_Xenomai_gpioirqbench

Wolfgang.

Re: [Xenomai-help] Handling interrupt - jitter

[Bosko Radivojevic]

On Fri, Apr 18, 2008 at 4:48 PM, Gilles Chanteperdrix
<gilles.chanteperdrix@xenomai.org> wrote:

>  It's funny to notice how much you asked for the 2.6.24 patch for
>  actually not using it.

BTW, 2.6.25 is out... :)

- Bosko

Re: [Xenomai-help] Handling interrupt - jitter

[Gilles Chanteperdrix]

On Mon, Apr 21, 2008 at 12:07 PM, Bosko Radivojevic
<bosko.radivojevic@domain.hid> wrote:
> On Fri, Apr 18, 2008 at 4:48 PM, Gilles Chanteperdrix
>  <gilles.chanteperdrix@xenomai.org> wrote:
>
>
> >  It's funny to notice how much you asked for the 2.6.24 patch for
>  >  actually not using it.
>
>  BTW, 2.6.25 is out... :)

I have just tried to apply the 2.6.24 patch to a 2.6.25 kernel there
does not seem to be much rejects, and they all seem easy to solve.
Anyway, I wait for Philippe to issue the x86 patch (and update the
I-pipe git noarch branch), and maybe for the AT91 people to issue
their patch (the I-pipe patch for ARM is such that you can apply it
with or without the AT91 patch, this requires some work).

-- 
 Gilles

Re: [Xenomai-help] Handling interrupt - jitter

[Philippe Gerum]

Gilles Chanteperdrix wrote:
> On Mon, Apr 21, 2008 at 12:07 PM, Bosko Radivojevic
> <bosko.radivojevic@domain.hid> wrote:
>> On Fri, Apr 18, 2008 at 4:48 PM, Gilles Chanteperdrix
>>  <gilles.chanteperdrix@xenomai.org> wrote:
>>
>>
>>>  It's funny to notice how much you asked for the 2.6.24 patch for
>>  >  actually not using it.
>>
>>  BTW, 2.6.25 is out... :)
> 
> I have just tried to apply the 2.6.24 patch to a 2.6.25 kernel there
> does not seem to be much rejects, and they all seem easy to solve.
> Anyway, I wait for Philippe to issue the x86 patch (and update the
> I-pipe git noarch branch)

I have nothing pending right now for 2.6.24-noarch. The public repo should be in
sync.

, and maybe for the AT91 people to issue
> their patch (the I-pipe patch for ARM is such that you can apply it
> with or without the AT91 patch, this requires some work).
> 


-- 
Philippe.

Re: [Xenomai-help] Handling interrupt - jitter

[Gilles Chanteperdrix]

On Mon, Apr 21, 2008 at 4:36 PM, Philippe Gerum
<philippe.gerum@domain.hid> wrote:
> Gilles Chanteperdrix wrote:
>  > On Mon, Apr 21, 2008 at 12:07 PM, Bosko Radivojevic
>  > <bosko.radivojevic@domain.hid> wrote:
>  >> On Fri, Apr 18, 2008 at 4:48 PM, Gilles Chanteperdrix
>  >>  <gilles.chanteperdrix@xenomai.org> wrote:
>  >>
>  >>
>  >>>  It's funny to notice how much you asked for the 2.6.24 patch for
>  >>  >  actually not using it.
>  >>
>  >>  BTW, 2.6.25 is out... :)
>  >
>  > I have just tried to apply the 2.6.24 patch to a 2.6.25 kernel there
>  > does not seem to be much rejects, and they all seem easy to solve.
>  > Anyway, I wait for Philippe to issue the x86 patch (and update the
>  > I-pipe git noarch branch)
>
>  I have nothing pending right now for 2.6.24-noarch. The public repo should be in
>  sync.

I meant the 2.6.25-noarch...

-- 
 Gilles

Re: [Xenomai-help] Handling interrupt - jitter

[Philippe Gerum]

Gilles Chanteperdrix wrote:
> On Mon, Apr 21, 2008 at 4:36 PM, Philippe Gerum
> <philippe.gerum@domain.hid> wrote:
>> Gilles Chanteperdrix wrote:
>>  > On Mon, Apr 21, 2008 at 12:07 PM, Bosko Radivojevic
>>  > <bosko.radivojevic@domain.hid> wrote:
>>  >> On Fri, Apr 18, 2008 at 4:48 PM, Gilles Chanteperdrix
>>  >>  <gilles.chanteperdrix@xenomai.org> wrote:
>>  >>
>>  >>
>>  >>>  It's funny to notice how much you asked for the 2.6.24 patch for
>>  >>  >  actually not using it.
>>  >>
>>  >>  BTW, 2.6.25 is out... :)
>>  >
>>  > I have just tried to apply the 2.6.24 patch to a 2.6.25 kernel there
>>  > does not seem to be much rejects, and they all seem easy to solve.
>>  > Anyway, I wait for Philippe to issue the x86 patch (and update the
>>  > I-pipe git noarch branch)
>>
>>  I have nothing pending right now for 2.6.24-noarch. The public repo should be in
>>  sync.
> 
> I meant the 2.6.25-noarch...
> 

Ok. I will probably address this in a couple of weeks.

-- 
Philippe.